How shared AI inspiration could reshape human creativity – and the machines that learn from it.
Published: September 2026 • Reading time: 9 min
Imagine the Web twenty years from now.
A researcher asks an LLM to help write a paper. An engineer asks one to prepare documentation. A student uses one to answer a question. A journalist uses one to structure an article. A programmer asks one to explain a bug. None of them simply copies what the machine produces. They read it, correct it, argue with it, perhaps rewrite half of it. The resulting material may be excellent. But it does not disappear when they are finished with it. It becomes part of the Web: Tomorrow’s papers, documentation, tutorials, articles and discussions. And some of that material will eventually help train the next generation of machines. Now multiply this process by hundreds of millions of people. We are beginning to create a loop:
There is nothing necessarily alarming about this. Human culture has always been recursive. We read other people, absorb their ideas and write new things. Scientists build on earlier scientists. Writers imitate writers before finding their own voices. But generative AI changes the scale – and perhaps the structure – of that recursion.
The first generation of large language models inherited something historically extraordinary: an enormous corpus produced mostly without large language models. Scientists wrote papers. Programmers argued late into the night on forums. Someone maintained a webpage devoted to an obscure radio receiver. Students asked questions that experts would never have thought to ask. Journalists interviewed people whose experiences existed nowhere else. Hobbyists documented failed experiments. People wrote beautifully and badly, carefully and impulsively. They disagreed, misunderstood each other, invented terminology and approached the same problems from different directions. The Web was noisy, repetitive and frequently wrong. But its messiness contained something precious: diversity of origin.
What happens when an increasing fraction of that diversity passes through the same generative machinery?
This is no longer an entirely hypothetical concern. Recent experiments have found that AI-assisted or AI-generated creative work can become more homogeneous at the collective level. In one study of 2,200 college-admissions essays, each additional human-written essay contributed more new ideas to the collective pool than an additional GPT-4 essay. Remarkably, increasing the diversity of individual AI outputs did not eliminate the gap in collective diversity.[1] A recent meta-analysis covering 19 studies and 61 effect sizes reaches a more cautious but similar conclusion: generative AI is associated with a small but significant homogenization effect in human–AI co-creation, although the magnitude depends on the task and on how people interact with the system.[2]
The danger, then, is not necessarily that AI-generated information becomes bad. Something subtler can happen:
An Information-Theoretic Limit
Information theory has a wonderfully unforgiving principle for what happens when information is repeatedly processed. Suppose information passes through a sequence
If Z receives its information about X only through Y, the Data Processing Inequality tells us that
In plain language: processing can reorganize information, compress it, reveal patterns in it and make it vastly more useful. But it cannot recreate distinctions about the original source that have already been lost along the way. The principle grew out of the information-theoretic framework established by Claude Shannon and is now one of the field’s basic results. Cover and Thomas give its standard treatment in Elements of Information Theory.[3]
Now imagine an idealized chain:
Here H is our accumulated human corpus, M_t a generation of models and S_t the material those models help us produce. This is an analogy, not a literal Markov chain. AI can produce genuinely new results. A mathematical system can derive a proof nobody has previously written down. A model can discover an algorithm, find a surprising strategy or combine known ideas into something its creators had never considered. Nothing in the Data Processing Inequality says otherwise. But the inequality suggests a different question:
When information repeatedly passes through similar representations, which distinctions survive, and which quietly disappear?
That question becomes more interesting once the loop closes. Research on so-called model collapse has already demonstrated one version of the problem inside machine learning. When models are recursively trained on model-generated data, information about the tails of the original distribution can progressively disappear.[4]
The human–AI ecosystem is vastly richer than that experiment. Humans contribute knowledge, judgment and new information at every stage. But this raises the central question of this essay:
Are we injecting enough independent variation back into the loop?
Generative abundance should not be confused with informational diversity.
One Hundred Authors
Imagine a simple experiment. We put one hundred knowledgeable people in one hundred rooms and give them the same difficult question. One person remembers a paper she read ten years ago. Another approaches the problem geometrically because that is how he learned to think. Someone misunderstands the question and, by accident, finds an interesting interpretation of it. One answer is brilliant. Several are pedestrian. A few are simply wrong.
Now repeat the experiment, except this time there is a state-of-the-art language model in every room. Within seconds, each screen fills with an articulate answer. The humans are not passive. They correct mistakes, remove weak arguments and add their own expertise. The second set of answers may well be better. But the hundred authors have acquired a hidden common collaborator. Their texts may still look different, yet their starting points are now more correlated. Similar arguments appear near the top. Similar examples suggest themselves. Some ways of framing the problem are repeatedly offered; others never appear.
A rare idea does not have to be censored to disappear. It only has to stop being suggested.
This is close to what the recent empirical literature is beginning to observe. The effect is not absolute—AI does not make everybody identical—but at scale, shared generative assistance can shift a collection of outputs toward greater similarity.[1,2,5]
And those outputs do not end with their authors. They become papers, webpages, documentation and discussions. Some may eventually enter future training corpora. What was already probable becomes slightly more visible. What was unusual becomes slightly harder to encounter. The loop closes:
Correlated Inspiration
There is an obvious objection. Humans have never created from nothing. Music grows from earlier music, mathematics from earlier mathematics, painting from earlier painting. Inspiration is itself a form of inheritance. But historically that inheritance has followed many paths. Different teachers, books, cultures, conversations and accidents shaped different minds. Human culture worked, in a loose sense, like an enormous mixture of experts: overlapping, but never quite seeing the world in the same way.
LLMs were trained on the traces left by this diversity. Now this enormous mixture increasingly turns to the same few models for inspiration. Why should that matter?
Because discovery is partly a search problem.
When many minds approach a question differently, they explore different parts of the space of possible ideas. Most paths lead nowhere. But occasionally the odd path—the one nobody else considered promising—is precisely where something new is found. A common AI assistant can make every explorer individually better while nudging many of them toward the same promising regions.
That creates a paradox: better individual search, but potentially narrower collective exploration.
History occasionally rewards intellectual outliers: an Archimedes or a Leonardo da Vinci, minds that followed unusual combinations of interests and lines of thought. We cannot know whether such figures would have thought differently in an AI-mediated world. That uncertainty is precisely the point. We do not know beforehand which unusual intellectual trajectory will matter.
The value of intellectual diversity is not that every different idea deserves to survive. Most do not. Its value is that we cannot know beforehand which unusual idea, method or question will turn out to matter. The problem is not inspiration.
It is correlated inspiration.
So What Should the Human Do?
The obvious response is more human supervision. That is necessary, but it is not sufficient. A human can carefully verify an AI-generated article while contributing very little outside the model’s original trajectory. We can correct its facts, improve its prose and choose the strongest of five arguments it proposes. That makes us good editors.
But the information ecosystem needs us to be something else as well: sources.
There is a meaningful difference between asking: “Give me five interesting research questions about this problem.”
and arriving with: “Something bothers me about the way we formulate this problem. What happens if we look at it this other way?”
In both cases AI can be enormously useful. But the intellectual direction begins in a different place. In the first case, the model proposes possible directions and the human selects from them. In the second, the human introduces a direction. The machine can then do what it does extraordinarily well: search, challenge the idea, find related work, derive consequences, generate counterexamples and expose weaknesses.
The point is not to keep AI at arm’s length. Quite the opposite. Come to the model with something.
A question. An observation. An objection. A strange analogy. Something you noticed at work. A result that does not fit. An idea that may turn out to be wrong. Then use the machine aggressively. Human–AI interaction does not reduce intellectual diversity simply because AI participates. What matters is whether humans remain active sources within that interaction rather than becoming only selectors of machine-generated possibilities.
To Ask, We Still Need to Learn
There is one catch: to arrive with a good question, we usually need to know something. A researcher questions an assumption because she understands why it was introduced. An engineer notices an anomaly because he knows what normal looks like. We connect two ideas because both were already somewhere in our minds.
Knowledge gives us the landscape against which something can appear surprising.
This is why education cannot simply outsource knowledge to AI. Students still need enough understanding to disagree, wonder and ask their own questions. The objective is not to know everything. It is to know enough to think independently:
AI can then take those directions much farther than we could alone.
Beyond Russell’s Limit
Bertrand Russell worried about a growing asymmetry: human knowledge expands, while the amount an individual can assimilate remains limited. In a previous essay, The AI Telco Engineer Against Russell’s Limit, I explored what this means in the age of generative AI through the idea of an “audit threshold”: if we cannot personally reproduce everything machines can produce, we must at least understand enough to interrogate and judge it.[6]
There is another side to that argument.
We must also understand enough to remain a source of questions rather than merely a consumer of answers. The answer is not to use AI less. It is to use it actively.
Let machines search farther than we can search, calculate faster than we can calculate and explore more alternatives than we could explore in a lifetime. But preserve enough knowledge and understanding to occasionally look at what the machine gives us and say: No. That is not quite the question I wanted to ask.
Perhaps that small act of intellectual independence will become more valuable, not less, as AI becomes more capable.
The first generation of large language models inherited an extraordinary archive of human intellectual diversity. Our responsibility is not merely to preserve it. It is to make sure that we still have something of our own to add.
References
[1] K. Moon, A. E. Green, and K. Kushlev, “Homogenizing effect of large language models (LLMs) on creative diversity: An empirical comparison of human and ChatGPT writing,” Computers in Human Behavior: Artificial Humans, vol. 6, 100207, 2025.
[2] A. de Rooij and M. M. Biskjaer, “Does generative AI make us think alike? A systematic review and meta-analysis of homogenisation effects in human–AI co-creation,” Behaviour & Information Technology, 2026.
[3] T. M. Cover and J. A. Thomas, Elements of Information Theory, 2nd ed., Wiley, 2006.
[4] I. Shumailov et al., “AI models collapse when trained on recursively generated data,” Nature, vol. 631, pp. 755–759, 2024.
[5] Z. Sourati, A. S. Ziabari, M. Dehghani, “The homogenizing effect of large language models on human expression and thought,” Trends in Cognitive Sciences, 2026.
[6] A. Giovanidis, “The AI Telco Engineer Against Russell’s Limit,” anastasiosgiovanidis.net, 2026.
Cite as
@misc{giovanidis2026dpi,
author = {Anastasios Giovanidis},
title = {When Everyone Asks the Same Machine},
year = {2026},
howpublished = {\url{https://anastasiosgiovanidis.net}},
note = {Online article}
}
Giovanidis, A. (2026). When Everyone Asks the Same Machine. anastasiosgiovanidis.net.